We hypothesise the study of acute protein perturbation in signal transduction by targeted anticancer drugs can predict drug sensitivity of these agents used as single agents and rational combination therapy. We assayed dynamic changes in 52 phosphoproteins caused by acute exposure (1hr) to clinically-relevant concentrations of 7 targeted anticancer drugs in 35 non small-cell lung cancer (NSCLC) cell lines and 16 samples of NSCLC cells isolated from patient pleural effusions. We studied drug sensitivities across 35 cell lines and synergy of combinations of all drugs in six cell lines (252 combinations). We developed orthogonal machine-learning approaches to predict drug response and rational combination therapy. Our methods predicted the most and least sensitive quartiles of drug sensitivity with an AUC of 0.79 and 0.78 respectively, while predictions based on mutations in three genes commonly known to predict response to the drug studied e.g. EGFR, PIK3CA and KRAS, did not predict sensitivity (AUC 0.5 across all quartiles). The machine-learning predictions of combinations was compared to experimentally-generated data showed a bias to the highest quartile of Bliss synergy scores, p=0.0243. We confirmed feasibility of running such assays on 16 patient samples of freshly isolated NSCLC cells from pleural effusions. We have provided proof of concept for novel methods of using acute ex-vivo exposure of cancer cells to targeted anticancer drugs to predict response as single agents or combinations. These approaches could compliment current approaches using gene mutations/amplifications/rearrangements as biomarkers, and demonstrate the utility of proteomics data to inform treatment selection in the clinic.